Spontaneous and driven cortical activity: implications for computation

Abstract

The traditional view of spontaneous neural activity as 'noise' has been challenged by recent findings suggesting that: (a) spontaneous activity in cortical populations is highly structured in both space and time, (b) the spatio-temporal structure of spontaneous activity is linked to the underlying connectivity of the cortical network, (c) spontaneous cortical activity interacts with external stimulation to generate responses to the individual presentations of a stimulus, (d) network connectivity is shaped in part by the statistics of natural signals and (e) ongoing cortical activity represents a continuous top-down prediction/expectation signal that interacts with incoming input to generate an updated representation of the world. These results can be integrated to provide a new framework for the study of cortical computation.

Fig 1

Visual summary of main concepts. (a) Coding of orientation by a population of V1 cells. Different oriented gratings produce a profile of activity centered at a different location. Each panel shows the activity of cells with different preferred orientations. (b) Assuming symmetry, the resulting manifold of population activity is a circle. (c) Conceptual view of how intrinsic dynamics, feedforward and feedback drive, combine to drive the cortical state. In each panel, a “force field” is shown that qualitatively describes how each of the individual mechanisms would act when the cortex is found in different states (in this case represented by the plane). The red dots indicate the equilibrium points for the feedforward, feedback and combined components. (d) A more general view of the framework. At any point in time primary visual cortex integrates feed-forward information (indicated by the variable Y) with top-down contextual information provided by extra-striate areas (denoted by the variable Z). The state of population activity in primary visual cortex is constrained to the manifold of natural signals, represented by the blue surface. One way of thinking about the integration of information is via Bayesian inference, where the V1 circuitry works to find the state within the manifold that maximizes the conditional probability, p (X | Y, Z). Under some assumptions [46], p (X | Y, Z) ~ p (X | Y) p (X | Z). A similar arrangement can be postulated at different levels in the visual hierarchy.